In this study, we used plasma samples from a longitudinal study analyzed on an aptamer-based 5000-plex proteomics as a discovery cohort to identify biomarker signatures and single biomarker candidates for LOPE. The most promising biomarker candidates, FAAH2, sFLT1, IL17RC and HTRA1, were differentially abundant between LOPE participants and healthy controls, and these proteins made up the highest performing biomarker signature at the last visit (visit 3). This visit was on average 7.6 weeks before the LOPE women gave birth, indicating that the samples were collected well before the onset of LOPE. Hence, the proposed biomarkers demonstrated the ability to differentiate between healthy and LOPE participants prior to clinical signs of PE. Furthermore, less than half of the LOPE women (45%) had developed hypertension at a later visit (mean GW 37, SD 0.7) that were not part of the current study (data not shown).Validation in the validation cohort showed that the predictive performance of our suggested biomarker signature at visit 3 (FAAH2, sFLT1, HTRA1, IL17RC, maternal BMI and age, and nulliparity) were reproducible (AUC 0.90 (95% CI 0.84–0.90) for LOPE and 0.95 (95% CI 0.91–0.99) for both EOPE and LOPE). We also found that the proteins contributed to the high prediction performance, as models including the clinical variables alone showed a decreased performance (Supplementary File 1: Fig. S2). The prediction performance increased from AUC = 0.9 (95% CI 0.84–0.90) (only including LOPE) to AUC = 0.95 (95% CI 0.91–0.99) with inclusion of both EOPE and LOPE in the PE group (Supplementary File 1: Fig. S1). Thus, the increase in AUC when including the EOPE cases in the validation cohort could indicate that the predictive proteins are linked to disease severity because EOPE is typically more severe than LOPE. Also, the findings make it plausible that our biomarker signature captures the complexity of PE and is robust across variants of PE. The protein abundances in the validation cohort were de-trended (adjusted for GW), so the increased prediction performance is not simply due to differences in GW between proteins in the validation cohort. However, an important difference to note is that the women in the PE group (both subtypes) in the validation cohort were already diagnosed with PE at the time of sampling, as opposed to the LOPE women in the discovery cohort. Thus, the validation illustrated that the biomarker signatures we identified before PE diagnosis were still able to separate groups in a cohort where the diagnosis had already been determined.The prediction performance of our suggested biomarker signature was comparable to the sFLT1/PGF ratio plus clinical variables within the discovery cohort and performed better at the third visit, i.e., with increased true positive rates, when the false positive rate was controlled at reasonably low values. The sFLT1/PGF ratio with the clinical variables performed better (AUC 0.80 (95% CI 0.66–0.85)) than our biomarker signature (AUC 0.76 (95% CI 0.61–0.82)) at visit 1 (Fig. 3d-e). At visit 2, the sFLT1/PGF ratio with the clinical variables performed poorer (AUC 0.80 (95% CI 0.65–0.85)) than our biomarker signature (AUC 0.86 (95% CI 0.78–0.94)). Lastly, the sFLT1/PGF ratio with the clinical variables (AUC = 0.87, 95% CI 0.80–0.94) performed equally well as our biomarker signature at visit 3 (AUC 0.88, 95% CI 0.80–0.96). By looking at the 95% CI, we see that the differences in model prediction performances measured in AUC are not statistically significantly different. However, for small values of false positive rate (FPR < 0.3) at visit 3, our biomarker signature resulted in increased true positive rates than the sFLT1/PGF ratio with the clinical variables. These comparisons indicated that our proposed biomarkers may have predictive abilities comparable to, or even superior to, the already established biomarker candidates. Thus, we believe that these new suggested biomarkers should be further investigated by the research community, as preeclampsia is a complicated and challenging syndrome to predict. It has been stated that the prediction performance of biomarkers may improve if novel biomarker candidates are combined with proteins already associated with PE and placental dysfunction, such as sFLT1, PGF and the sFLT1/PGF ratio7. However, in our case, the sFLT1/PGF ratio did not seem to add to the predictive performance of our selected biomarker signature, as our biomarker signature with the sFLT1/PGF ratio (Fig. 3d-f) demonstrated relatively similar performance per visit as our biomarker signature. The sFLT1/PGF ratio in the present study was measured by an aptamers-based technology rarely used in clinical settings and not by commercially available kits, such as ELISA, which is more common in the clinic. However, we have previously reported a high correlation between the aptamers for sFLT1 and PGF and ELISA antibodies (Spearman r > 0.75, p < 0.05)39.According to the “Tissue-based map of the human proteome”40, sFLT1, FAAH1, HTRA and IL17RC originate from multiple organ tissues, but mainly from the placenta, pancreas, brain and liver, respectively (Supplementary File 1: Fig. S3). We have recently described the release of FAAH2, HtrA1 and sFLT1 from the placenta into the maternal circulation in healthy pregnancies18. The placental origin of these biomarker candidates fits well with placental dysfunction being a key feature of PE10. FAAH2 is one of two degrading membrane-bound hydrolases in the endocannabinoid system41. The endocannabinoid system has an important impact on a wide range of fertility and pregnancy outcomes42,43. FAAH2 degrades bioactive lipids such as anandamide (N-arachidonoylethanolamine), an endogenously produced cannabinoid-like lipid mediator that binds to the cannabinoid receptors 1 and 2. When comparing PE with healthy pregnancies, Aban et al.44 reported reduced FAAH activity in placental villi45 in PE pregnancies, whereas Fugedi et al.46 reported no differences in FAAH expression nor immunoreactivity in the placenta47. To the best of our knowledge, no data has previously been published on FAAH2 specifically in relation to PE.HTRA1 is a secreted protein with an important role in degradation of misfolded proteins as well as cell growth, including the development of organs such as the placenta48. The expression of this protein in placental trophoblasts increases with gestation48. Several studies have demonstrated abnormal levels of HtrA1 in placental tissue and maternal serum in PE49,50,51,52,53, whereas other studies did not54. Our data support an elevated abundance of HTRA1 in the circulation at 28–34 GW in women with PE.The observed elevated abundance of sFLT1 at GW 28–34 among the PE women is in strong accordance with former research2,55,56, and elevated placental release of sFLT1 to the maternal circulation in preeclamptic pregnancies has previously been shown by our group2. sFLT1 consists of the extracellular ligand-binding domain of vascular endothelial growth factor receptor 1 (VEGFR1) and can therefore act as an antagonist of the angiogenic factors Vascular endothelial growth factor (VEGF) and PGF by hindering their binding to their respective receptors57.The inflammatory cytokine interleukin-17 has been implicated in the pathogenesis of PE and suggested as a biomarker candidate58. A soluble, secreted form of IL17RC, lacking transmembrane and intracellular domains, may function as an extracellular antagonist of cytokine signaling and IL-17 specifically59. A study of a rat preeclampsia model (RUPP rats) has indicated that IL-17RC can reduce mean arterial pressure, placental levels of TNF-α and macrophage inflammatory protein-3 α, as well as normalize placental cytolytic natural killer cell levels59. In humans, the frequency of IL-17-expressing CD4 + T cells decreased with gestation in healthy pregnancies but remained unchanged in preeclampsia60, supporting a shift towards higher IL-17 production in PE as compared to healthy pregnant women61. We speculate that higher IL-17RC levels in PE may be a compensatory mechanism aiming to reduce the effects of placental ischemia.Comparing our differentially abundant proteins to previous studies indicates that we have identified novel biomarker candidates for LOPE. A systematic review by Navajas et al.8 presented 559 protein biomarker candidates with differential abundance between controls and PE (EOPE and/or LOPE or type not specified) in plasma and serum samples analyzed by mass spectrometry (MS) and/or microarray8. Among the 58 differentially abundant proteins at visit three in our results, 14 proteins overlapped with Navajas’ findings (Fig. 7), underpinning their relevance as potential biomarkers. Compared to MS and small-scale microarrays, the larger 5000-plex microarray in this study may have contributed to the identification of more biomarker candidates for LOPE than in any previous study. Thus, the 44 differentially abundant proteins not overlapping with proteins presented by Navajas et al.8 are interesting due to their novelty as biomarker candidates. In addition to the use of a large-scale proteomics platform, differences in demographic characteristics between the present and former studies possibly impacted the findings. A study of LOPE by Erez et al.12 resembled the present study regarding longitudinal design, study population (controls vs. LOPE) and proteomics platform (SomaScan), but with a lower amount of quantified proteins (1125 proteins). Erez’ study identified 36 differentially abundant proteins of which only three, SIGLEC6, NID1 and HSPA1A, overlapped with the 58 differentially abundant proteins identified in the present study. Despite the differences in the number of quantified proteins, it is somewhat surprising that only three differentially abundant proteins overlapped. Possible explanations may be younger study participants (both PE and controls) (median 23 vs. 31.5), more often smokers (18.5% vs. 1%), higher BMI (median 28.3 vs. 24.5), less nulliparous women (35% vs. 62%) and mostly African Americans in Erez et al.’s study, while the present study contained only Caucasians. The three proteins that still overlap may be universal biomarker candidates.Fig. 7Overlap between the review of Navajas et al.8 and the current study. Proteins as biomarker candidates of PE in any of the included studies (Navajas), proteins marked as biomarker candidates in third trimester (Navajas 3 trim) and the differentially abundant proteins in the current study: the 58 unique Entrez Gene symbol names (Current study 3 trim LOPE). Proteins marked as biomarker candidates in third trimester were mostly (96%) markers of PE in general (according to Navajas et al.8), not specifically for LOPE (2%) or EOPE (1%).Among the differentially abundant proteins, the biological processes “response to BMP”, “BMP signaling pathway”, “cellular response to BMP stimulus”, “regulation of BMP signaling pathway” and “negative regulation of cellular response to growth factor stimulus”, as well as the molecular function ontology “BMP binding” (Fig. 6) were overrepresented. BMPs are proteins that “constitute the largest subdivision of the transforming growth factor-β family of ligands”, and the BMP-related processes are e.g., embryonic patterning, organogenesis, angiogenesis, vascular integrity and inflammation62. Even though 12 BMPs (BMPR1B, BMP15, BMPER, BMP2, BMP4, BMP8B, BMP3, BMP1, BMP10, BMPR2, BMP4, BMP6) were included in our proteomics platform (Additional File 2: Table S3 in18), neither were differentially abundant between LOPE and controls at any visit (Supplementary File 2: Table S1). The proteins associated with these BMP-related ontologies (Supplementary File 2: Table S3) showed higher abundance in LOPE than controls. In a similar study with more samples, but fewer quantified proteins, BMP1 was identified as a predictor for LOPE early in pregnancy (8–16 and 16–33 GW), but not later (> 22.1 GW)12. These findings suggest that even though BMPs themselves were not significantly differential, proteins associated with BMP processes and functions might serve as universal biomarker candidates for LOPE across cohorts.Since the proposed PE biomarkers performed best at predicting LOPE in the third trimester, we suggest that the biomarker candidates identified in this study should be considered for triage and risk stratification rather than early prediction/screening for PE. In this way, the suggested biomarker candidates can be used to identify women considered at risk of LOPE or women with symptoms of uncertain significance. Hence, the biomarkers may differentiate between women in need of closer monitoring and women who are at low risk and may return to regular prenatal care. Third trimester triage biomarker candidates will be clinically valuable, given that PE is most prevalent close to term pregnancy1,7. However, this must be validated, see next the paragraphs.Although the current study showed promising results, the suggested biomarker candidates need to be validated in a cohort where negative predictive values can be validly calculated. In the present study, we have selected a subset of the whole STORK cohort as controls for our discovery cohort and included all available LOPE cases. As the number of participants in the control and PE groups does not correspond to the true prevalence of PE, we cannot calculate a valid negative predictive value.The increased prediction performance of the EN and ENSS algorithms relative to RF may be explained in part by the fact that the known risk factors (age, BMI and nulliparity) were always included as covariates in the models during cross-validation runs, while for RF such enforcement of covariates was not possible, leaving the decision of inclusion to the RF algorithm.The discovery dataset was combined with the validation dataset to de-trend the validation dataset prior to validation. We are aware that this step created a dependency between the validation dataset and the discovery dataset that may have influenced the validation results. However, we still consider this step crucial because the discovery cohort controls were the only controls sampled at GWs that matched PE who were sampled before GW 38.Using the 5000-plex Somalogic platform, we were able to quantify ¼ of the total number of proteins circulating in the human body40. However, the present study was able to quantify a significantly higher number of proteins than previous studies8,12.The current criteria for diagnosing PE have changed since participants were recruited into the discovery cohort, and proteinuria is no longer a necessary diagnostic criterion for PE63. Hence, some of our control women may belong to the PE group according to novel diagnostic criteria.The reproducibility of protein abundances quantified by SomaScan to other proteomics technologies (ELISA, Mass spectrometry, Olink) is highly dependent on the single protein64,65. Furthermore, SOMAmer kits may be potential novel clinical tools for prediction of PE, as aptamers are already, to some extent, in clinical use and several are under clinical testing66. Since the SomaScan quantifies many more proteins than other platforms, it is well suited for discovery studies. However, all findings need further validation, both on other platforms and in independent cohorts. If the biomarkers are proven useful in further validation studies, ELISA antibodies or other clinically useful measurement methods may be developed.
